Motivation
In recent years, urban floods, exacerbated by climate change and urbanization, demand accurate mapping for effective disaster management. Synthetic Aperture Radar (SAR) offers all-weather capabilities, but current SAR-based flood mapping lacks exploration of deep learning methods. Traditional approaches and supervised deep learning dominate, with limited focus on semi-supervised and unsupervised methods. Ground truth acquisition for urban flood mapping is challenging and time consuming. This project aims to address these gaps, investigating the potential of semi-supervised and unsupervised learning in SAR-based urban flood mapping. By comparing these techniques with traditional methods, the project seeks to enhance accuracy and timeliness in disaster management efforts.

Goals
Propose a semi-supervise learning based methodology for large-scale urban flood mapping Requirements for the students.
- Experience with one of the established, selected DL frameworks (e.g., PyTorch)
- Basic programming skills in Python
- Basic understanding of remote sensing image semantic segmentation algorithm

Apply to this project here